diff mbox series

[2/3] tune-cortexa32: enable Thumb-2 and fix AArch32 crypto FPU selection

Message ID 20260701-tune-a32-v1-2-a43f28b1d7a4@gmail.com
State New
Headers show
Series tune-cortexa32: AArch32 Thumb-2/crypto fixes, plus a QEMU machine | expand

Commit Message

Alex Kiernan July 1, 2026, 11:58 a.m. UTC 
Cortex-A32 is an AArch32-only ARMv8-A core, but its tune was configured
as if it were a plain 64-bit-style tune: TUNE_FEATURES lacked both 'arm'
and 'thumb', so userspace and the kernel were built as fixed-width ARM
rather than Thumb-2, and the package architecture names did not match a
Thumb-2 build.

Add 'thumb' so the core is built as Thumb-2, and 'arm' so that
feature-arm-thumb.inc honours a recipe's explicit ARM_INSTRUCTION_SET =
"arm" instead of silently forcing it to thumb. Without 'arm', ARM_M_OPT
is pinned to "thumb" and recipes that request ARM warn, e.g.:

  Recipe 'libmad' selects ARM_INSTRUCTION_SET to be 'arm', but tune
  configuration overrides it to 'thumb'

On AArch32 the ARMv8 crypto extensions (AES/SHA) are part of the FPU,
not just the -march ISA. The 'crypto' feature only appends '+crypto' to
-march and leaves -mfpu at plain 'neon', so building the AES intrinsics
(e.g. mbedtls aesce.c) fails:

  error: inlining failed in call to 'always_inline' 'vaesdq_u8':
  target specific option mismatch

Select the crypto FPU by appending crypto-neon-fp-armv8 as the last word
of TUNE_CCARGS_MFPU when both 'crypto' and 'neon' are active, mirroring
the vfpv4+neon -> neon-vfpv4 idiom in feature-arm-neon.inc.

Update PACKAGE_EXTRA_ARCHS to match the resulting package arches:
cortexa32hf-neon becomes cortexa32t2hf-neon, and the crypto variant
becomes cortexa32t2hf-neon plus cortexa32t2hf-crypto-neon-fp-armv8.

AI-Generated: Claude Code (Claude Opus 4.8)
Signed-off-by: Alex Kiernan <alex.kiernan@gmail.com>
Signed-off-by: Alex Kiernan <alexk@a-squared-projects.uk>
---
 meta/conf/machine/include/arm/armv8a/tune-cortexa32.inc | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)
diff mbox series

Patch

diff --git a/meta/conf/machine/include/arm/armv8a/tune-cortexa32.inc b/meta/conf/machine/include/arm/armv8a/tune-cortexa32.inc
index 0eb938a2403a..f8f876813302 100644
--- a/meta/conf/machine/include/arm/armv8a/tune-cortexa32.inc
+++ b/meta/conf/machine/include/arm/armv8a/tune-cortexa32.inc
@@ -5,14 +5,22 @@  TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'cortexa32', ' -mcpu=corte
 
 require conf/machine/include/arm/arch-armv8a.inc
 
+# On AArch32 the ARMv8 crypto extensions (AES/SHA) live in the FPU, not just
+# -march. OE-Core's 'crypto' feature only adds '+crypto' to -march, leaving the
+# FPU at plain 'neon', so GCC refuses to inline the AES intrinsics used by e.g.
+# mbedtls' aesce.c ("target specific option mismatch"). Select the crypto FPU by
+# making it the last word of TUNE_CCARGS_MFPU - the same idiom OE-Core uses for
+# the vfpv4+neon -> neon-vfpv4 combination in feature-arm-neon.inc.
+TUNE_CCARGS_MFPU .= "${@bb.utils.contains('TUNE_FEATURES', [ 'crypto', 'neon' ], ' crypto-neon-fp-armv8', '', d)}"
+
 # Little Endian base configs
 AVAILTUNES += "cortexa32 cortexa32-crypto"
 ARMPKGARCH:tune-cortexa32             = "cortexa32"
 ARMPKGARCH:tune-cortexa32-crypto      = "cortexa32"
 # We do not want -march since -mcpu is added above to cover for it
-TUNE_FEATURES:tune-cortexa32          = "armv8a cortexa32 crc callconvention-hard neon"
+TUNE_FEATURES:tune-cortexa32          = "arm armv8a cortexa32 crc callconvention-hard neon thumb"
 TUNE_FEATURES:tune-cortexa32-crypto   = "${TUNE_FEATURES:tune-cortexa32} crypto"
-PACKAGE_EXTRA_ARCHS:tune-cortexa32             = "${PACKAGE_EXTRA_ARCHS:tune-armv8a-crc} cortexa32 cortexa32hf-neon"
-PACKAGE_EXTRA_ARCHS:tune-cortexa32-crypto      = "${PACKAGE_EXTRA_ARCHS:tune-armv8a-crc-crypto} cortexa32 cortexa32hf-neon cortexa32hf-neon-crypto"
+PACKAGE_EXTRA_ARCHS:tune-cortexa32             = "${PACKAGE_EXTRA_ARCHS:tune-armv8a-crc} cortexa32 cortexa32t2hf-neon"
+PACKAGE_EXTRA_ARCHS:tune-cortexa32-crypto      = "${PACKAGE_EXTRA_ARCHS:tune-armv8a-crc-crypto} cortexa32 cortexa32t2hf-neon cortexa32t2hf-crypto-neon-fp-armv8"
 BASE_LIB:tune-cortexa32               = "lib"
 BASE_LIB:tune-cortexa32-crypto        = "lib"